1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
use super::super::args::{TraceArgs, TraceSub};
use crate::exit_codes;
use assay_core::trace;
mod import_mcp;
pub async fn cmd_trace(args: TraceArgs, legacy_mode: bool) -> anyhow::Result<i32> {
match args.cmd {
TraceSub::Ingest { input, output } => {
trace::ingest::ingest_file(&input, &output)?;
eprintln!(
"Ingested trace from {} to {}",
input.display(),
output.display()
);
Ok(exit_codes::OK)
}
TraceSub::IngestOtel {
input,
db,
suite: _suite,
out_trace,
} => {
use assay_core::storage::Store;
use assay_core::trace::otel_ingest::{convert_spans_to_episodes, OtelSpan};
use std::io::BufRead;
let file = std::fs::File::open(&input)
.map_err(|e| anyhow::anyhow!("failed to open input file: {}", e))?;
let reader = std::io::BufReader::new(file);
let mut spans = Vec::new();
for line in reader.lines() {
let line = line?; // propagate error
if line.trim().is_empty() {
continue;
}
let span: OtelSpan = serde_json::from_str(&line)
.map_err(|e| anyhow::anyhow!("failed to parse OTel span: {}", e))?;
spans.push(span);
}
let events = convert_spans_to_episodes(spans);
let count = events.len();
let store = Store::open(&db)?;
store.init_schema()?; // ensure tables exist
store.insert_batch(&events, None, None)?;
eprintln!(
"Ingested {} OTel spans as {} V2 events into {}",
count,
events.len(),
db.display()
);
if let Some(out_path) = out_trace {
let f = std::fs::File::create(&out_path)
.map_err(|e| anyhow::anyhow!("failed to create output trace file: {}", e))?;
let mut writer = std::io::BufWriter::new(f);
for event in &events {
use std::io::Write;
let json = serde_json::to_string(event)?;
writeln!(writer, "{}", json)?;
}
eprintln!("Wrote trace replay file to {}", out_path.display());
}
Ok(exit_codes::OK)
}
TraceSub::Verify { trace, config } => {
let cfg = assay_core::config::load_config(&config, legacy_mode, false)
.map_err(|e| anyhow::anyhow!("failed to load config: {}", e))?;
trace::verify::verify_coverage(&trace, &cfg)?;
Ok(exit_codes::OK)
}
TraceSub::PrecomputeEmbeddings {
trace,
config,
embedder,
model,
output,
} => {
let cfg = assay_core::config::load_config(&config, legacy_mode, false)
.map_err(|e| anyhow::anyhow!("failed to load config: {}", e))?;
// Build embedder (simplified version of build_runner logic)
use assay_core::providers::embedder::{
fake::FakeEmbedder, openai::OpenAIEmbedder, Embedder,
};
use std::sync::Arc;
let embedder_client: Arc<dyn Embedder> = match embedder.as_str() {
"openai" => {
let key = std::env::var("OPENAI_API_KEY")
.map_err(|_| anyhow::anyhow!("OPENAI_API_KEY required for precompute"))?;
Arc::new(OpenAIEmbedder::new(model.clone(), key))
}
"fake" => Arc::new(FakeEmbedder::new(&model, vec![0.1; 1536])), // Mock vector
_ => anyhow::bail!("unknown embedder: {}", embedder),
};
let out_path = output.unwrap_or_else(|| trace.clone()); // Default overwrite logic? No, let's play safe.
// If output is None, maybe we should warn?
// The user args say output is Option.
// Let's assume overwrite if not provided, as per "enrichment" philosophy.
// Wait, I can't overwrite input while reading it easily without slurp.
// `precompute_embeddings` takes input path and output path.
// If output is None, we should likely write to a temp file and rename.
let final_output = out_path.clone();
let effective_output = if final_output == trace {
// If rewriting in place, write to temp first
let mut temp = trace.clone();
temp.set_extension("tmp.jsonl");
temp
} else {
final_output.clone()
};
trace::precompute::precompute_embeddings(
&trace,
&effective_output,
embedder_client,
&model,
&cfg,
)
.await?;
if effective_output != final_output {
std::fs::rename(&effective_output, &final_output)?;
}
println!(
"Precomputed embeddings for {} -> {}",
trace.display(),
final_output.display()
);
Ok(exit_codes::OK)
}
TraceSub::PrecomputeJudge {
trace,
config,
judge,
judge_model,
output,
} => {
let cfg = assay_core::config::load_config(&config, legacy_mode, false)
.map_err(|e| anyhow::anyhow!("failed to load config: {}", e))?;
// Build judge service
// This is complex, requires JudgeRuntimeConfig etc.
// We need to implement a lightweight builder or reuse existing one.
// Reusing `assay_core::judge::JudgeService::new` requires `JudgeStore` and `LlmClient`.
// We can use a NullStore or MemoryStore? Precompute usually implies we don't care about caching results IN the DB,
// but we might want to USE the validation logic.
// Actually, we are WRITING to the trace JSONL. We probably don't need the sqlite store enabled?
// `JudgeService` requires a `JudgeStore` trait object? It takes `JudgeCache` struct which takes `Store`.
// This dependency chain `JudgeService -> JudgeCache -> Store -> Sqlite` makes lightweight usage hard.
// Maybe we should bypass `JudgeService` and call `LlmClient` directly?
// BUT `JudgeService` encapsulates the prompt logic (`templates::render_judge_prompt`). We NEED that.
// So we need a Store. We can open an in-memory SQLite store?
use assay_core::storage::Store;
let store = Store::memory()?;
// We need to init schema for it to work?
store.init_schema()?;
let judge_store = assay_core::storage::judge_cache::JudgeCache::new(store);
let model = judge_model
.clone()
.unwrap_or_else(|| "gpt-4o-mini".to_string());
use assay_core::providers::llm::{fake::FakeClient, openai::OpenAIClient, LlmClient};
use std::sync::Arc;
let client: Option<Arc<dyn LlmClient>> = match judge.as_str() {
"openai" => {
let key = std::env::var("OPENAI_API_KEY")
.map_err(|_| anyhow::anyhow!("OPENAI_API_KEY required"))?;
Some(Arc::new(OpenAIClient::new(model.clone(), key, 0.0, 1000)))
}
"fake" => Some(Arc::new(FakeClient::new(model.clone()))),
_ => anyhow::bail!("unknown judge provider: {}", judge),
};
let config = assay_core::judge::JudgeRuntimeConfig {
enabled: true,
provider: judge.clone(),
model: Some(model),
samples: 1, // Precompute usually deterministic 1 pass?
temperature: 0.0,
max_tokens: 1000,
refresh: true,
reliability: Default::default(),
system_prompt_version: String::new(),
};
let service = assay_core::judge::JudgeService::new(config, judge_store, client);
let out_path = output.unwrap_or_else(|| trace.clone());
let final_output = out_path.clone();
let effective_output = if final_output == trace {
let mut temp = trace.clone();
temp.set_extension("tmp.jsonl");
temp
} else {
final_output.clone()
};
trace::precompute::precompute_judge(&trace, &effective_output, &service, &cfg).await?;
if effective_output != final_output {
std::fs::rename(&effective_output, &final_output)?;
}
println!(
"Precomputed judge scores for {} -> {}",
trace.display(),
final_output.display()
);
Ok(exit_codes::OK)
}
TraceSub::ImportMcp {
input,
out_trace,
format,
episode_id,
test_id,
prompt,
} => {
let format_enum = assay_core::mcp::McpInputFormat::from_cli_label(&format)
.ok_or_else(|| anyhow::anyhow!("unknown format: {}", format))?;
import_mcp::run(import_mcp::ImportMcpArgs {
input,
out_trace,
format: format_enum,
episode_id,
test_id,
prompt,
})?;
Ok(exit_codes::OK)
}
}
}